==================================================

Q:

I am a student from Japan. I am very interested in your excellent work titled "Optimization of robust loss functions for weakly-labeled image taxonomies", which is published in IJCV2012.

Currently, I was trying to run your shared code on Linux machine.  I already had the related feature data e.g. "tmp/julian/EMMCVPR/features". The code in file " graphdata.cpp" indicated that following files are needed to be loaded:

/media/disk-1/julian_data/features/desc_test_%d_dense_%d.fvecs   (I have)
/media/disk-1/julian_data/features/desc_test_%d_dense_%d.fvecs    (I have)
/media/disk-1/julian_data/features/changes_class_imagenetTrain    (I have)
/media/disk-1/julian_data/features/test.list                                        (I have)
/media/disk-1/julian_data/features/cost_matrix.txt                            (I have)

but the files

media/disk-1/julian_data/svm/svm_%d_%d.svm  (I don't have)

May I have question that how can I get the raw data about the initial ova-svm model parameter files (such as "media/disk-1/julian_data/svm/svm_%d_%d.svm") ?

I would be appreciated for your kind help!

A:

Sure thing, I've put a tar.gz of that data here for you:
http://jmcauley.ucsd.edu/data/EMMCVPR/features_and_svm.tar.gz

Q:

With your provided data, I can load the features and run the code. However, I met another running problem:

-----------------------------------------------------------------------------------------------------------------------
When I set the temp_train.conf with BMRM.lambda 1e-2, after the program running for a period,
it output the message: 
"[BMRM error message] (Thrower: CBMRM::ConfirmProgramParameters())
At least one of the BMRM.{epsilonTol, gammaTol, relEpsilonTol, relGammaTol} must be > 0 "

What do you think about this problem? Since I did not set any options for these items, e.g. epsilonTol, 
gammaTol, etc.
-----------------------------------------------------------------------------------------------------------------------

# Since I don't have the " configfiles/graphmatch.l2.empty.conf" file mentioned in batch file " experiment.sh",
so I use the "graphmatch.l2.conf" in bmrm2.1 package as a template file instead, and add some items (BMRM options) for training.

# Since i only have a small size GPU, so I uncomment the code lines using cublas functions in the "graphdata.cpp" and "graphmatchloss.cpp" files,
and then compile and run the code. 

I would be appreciated for your suggestions.

A:

The config file in question is attached.:

string Solver.type BMRM
string BMRM.innerSolverType L2N2_DaiFletcherPGM
int L2N2_DaiFletcherPGM.gradIdleAge 10
int L2N2_DaiFletcherPGM.maxGradSetSize 200
int Loss.verbosity 0
string Loss.lossFunctionType GRAPHMATCH

/////////////////////////////////////////////////////
// These are the parameters that may need changing //
/////////////////////////////////////////////////////

// tolerance for epsilon termination criterion (set negative value to disable this criterion)
double BMRM.epsilonTol 1e-2
// number of iterations before termination
int BMRM.maxNumOfIter 2500

string Model.modelFile model
string Model.previousModel model
string Model.hotStartModel model